991715ab76
- llm_extract.py: match mode now window-parallel with retrieval pre-filter,
claim dedup, retry, and enable_thinking=false (vLLM) -> ~36x faster per call;
n_failed_windows/ok flags so an interrupted run never records bogus 0s.
- build_rdf_dataset.py:
- gold now includes the share-class level (hasShareClass/ticker/className)
- grounding modes alias|llm|name|context|none (--grounding); llm reads the
role-check verdicts from match_all.jsonl
- label stage: per-triple extractable + per-sample FULL/PARTIAL/NONE
- trainset stage: combines GROUNDED triples with focused TEXT EXCERPTS cut
around the actual provider statement (evidence), not the multi-MB book
- split --src to split trainset.jsonl (trust-level, no leakage)
- helper scripts: watch_match.sh, resume_match.sh (crash/sleep-safe resume),
finalize_dataset.sh
- final dataset: 335/335 trusts, 85% text<->gold agreement, 334 samples,
10,689 grounded triples, train/val/test 264/35/35
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
67 lines
2.8 KiB
Bash
67 lines
2.8 KiB
Bash
#!/usr/bin/env bash
|
||
# ------------------------------------------------------------------
|
||
# watch_match.sh — Fortschritt des LLM-Match-Laufs ueberwachen
|
||
#
|
||
# bash watch_match.sh -> Dashboard, aktualisiert alle 15s (Strg-C beendet)
|
||
# bash watch_match.sh live -> rohes Log live mitlaufen (tail -f)
|
||
# bash watch_match.sh once -> einmaliger Stand, dann zurueck zur Shell
|
||
#
|
||
# Beendet nur die ANSICHT, nicht den Match-Lauf (der laeuft im Hintergrund).
|
||
# ------------------------------------------------------------------
|
||
LOG="/private/tmp/claude-501/-Users-florianherzog-development-AISE-Admin-Thesis--Florian-Code/745ff40e-0ecf-4430-886c-33fbc724df69/tasks/bfe2v693w.output"
|
||
OUT="data/rdf_poc/match_remaining.jsonl"
|
||
TOTAL=256
|
||
SERVER="http://silicon.fhgr.ch:7080/v1/models"
|
||
|
||
show_once () {
|
||
local done last errs srv tg tt
|
||
done=$(wc -l < "$OUT" 2>/dev/null | tr -d ' '); done=${done:-0}
|
||
last=$(grep -oE '\[[0-9]+/'"$TOTAL"'\]' "$LOG" 2>/dev/null | tail -1)
|
||
errs=$(grep -cE 'error|ERROR|Traceback' "$LOG" 2>/dev/null | tr -d ' ')
|
||
# Server erreichbar?
|
||
# 2 Versuche mit grosszuegigem Timeout: unter Last antwortet /v1/models
|
||
# langsamer, ein kurzes Timeout meldet faelschlich "nicht erreichbar".
|
||
if curl -s -m 20 "$SERVER" >/dev/null 2>&1 || curl -s -m 20 "$SERVER" >/dev/null 2>&1; then
|
||
srv="OK"; else srv="NICHT erreichbar (oder ausgelastet)"; fi
|
||
# bisherige Uebereinstimmung aus den fertigen Samples
|
||
read tg tt < <(python3 - "$OUT" <<'PY' 2>/dev/null
|
||
import json,sys
|
||
g=t=0
|
||
try:
|
||
for l in open(sys.argv[1]):
|
||
r=json.loads(l)
|
||
for x in r.get("triples",[]):
|
||
t+=1; g+=1 if x.get("llm_grounded") else 0
|
||
except FileNotFoundError: pass
|
||
print(g,t)
|
||
PY
|
||
)
|
||
tg=${tg:-0}; tt=${tt:-0}
|
||
local pct=0; [ "$tt" -gt 0 ] && pct=$((100*tg/tt))
|
||
local dpct=$((100*done/TOTAL))
|
||
printf "\n LLM-Match-Lauf\n"
|
||
printf " ------------------------------------------\n"
|
||
printf " Fortschritt : %d / %d Trusts (%d%%) letzte Logzeile: %s\n" "$done" "$TOTAL" "$dpct" "${last:-—}"
|
||
printf " Tripel : %d von %d belegt (%d%% Uebereinstimmung bisher)\n" "$tg" "$tt" "$pct"
|
||
printf " Fehler/Log : %s\n" "${errs:-0}"
|
||
printf " vLLM-Server : %s\n" "$srv"
|
||
printf " ------------------------------------------\n"
|
||
}
|
||
|
||
case "${1:-dash}" in
|
||
live) echo "Live-Log (Strg-C beendet die Ansicht):"; exec tail -f "$LOG" ;;
|
||
once) show_once ;;
|
||
*)
|
||
echo "Dashboard – aktualisiert alle 15s (Strg-C beendet die Ansicht, nicht den Lauf)"
|
||
while true; do
|
||
clear 2>/dev/null
|
||
show_once
|
||
done=$(wc -l < "$OUT" 2>/dev/null | tr -d ' ')
|
||
if grep -q "matched .* samples" "$LOG" 2>/dev/null; then
|
||
echo " FERTIG. $(grep 'matched .* samples' "$LOG" | tail -1)"
|
||
break
|
||
fi
|
||
sleep 15
|
||
done ;;
|
||
esac
|